What's New in the World of AI? We've Successfully Completed Another Phase of DigiHavel!

What's new in the world of AI? We've successfully completed another phase of DigiHavel! During the pilot, students and teachers posed thousands of unique and often very sophisticated questions. Thank you very much. Thanks to them, the next version will be even smarter. I continue my little lecture tour. The next one will be on 1st December at the University of Economics.

But the scientists at Meta AI (Facebook) have not been idle either, as they launched a demo of a large language model named Galactica in mid-November. The model is intended to store, combine, and reason about scientific knowledge. Its strong suit is said to be technical knowledge, such as equations in LaTeX, mathematics, and the like. The model was trained "on 106 billion tokens of scientific texts and publicly available data. This includes, in particular, articles, textbooks, scientific websites, encyclopaedias, knowledge databases, and so on. Just two days after the launch of the online demo, the public model was halted following a wave of criticism from some scientists and the media. The model proved to be a potentially dangerous generator of pseudoscientific nonsense. It is clear that there was a misunderstanding of how these models work today, and many laypeople had exaggerated expectations of this model. Similar to GPT-3, it fabricates realistically looking articles in which it authentically combines truth and fiction. Many people did not understand that the outputs of the model should be taken more as inspiration, tips, and recommendations, and it is advisable to verify them independently. I believe it can be a great tool if we know how to use it correctly. Just like with GPT-3, several models have been released here according to size, from mini with 125 million parameters to huge with 120 billion parameters.

I tried the Galactica standard model (even this intermediate model with 7 billion parameters takes up nearly 30 gigabytes) on DigiHavel, and it "recommended" collaboration with the National Programme for the Development of Artificial Intelligence in the Czech Republic (NP AI), as well as with CTU, CTU, and funding from the Ministry of Education, Youth and Sports. It suggests I consider using the Dialogflow framework. For specific scientific questions, the standard model responds in accordance with the paper worse than the largest GPT-3. It's a pity that it's no longer so trivial to try the huge model.

You’ve probably also noticed the new image generation model, Stable Diffusion v2. The original SD V1 model from CompVis was unique primarily for its open-source code and has created hundreds of different models and improvements around the world. What distinguishes version 1 from version 2? I particularly appreciated the resolution. The output can now be in the default resolution of 768×768 pixels. Stable Diffusion 2.0 now includes an Upscaler Diffusion model that quadruples the image resolution. Furthermore, it should make it more difficult for users to imitate the styles of specific artists or generate NSFW (nudity, pornography, etc.) outputs.

Sources: https://www.seznamzpravy.cz/clanek/tech-techmix-po-protestech-vedcu-meta-novou-ai-vypnula-dva-dny-od-spusteni-219927

https://paperswithcode.com/paper/galactica-a-large-language-model-for-science-1

https://www.youtube.com/watch?v=ZTs_mXwMCs8

SD2: https://huggingface.co/stabilityai/stable-diffusion-2/discussions https://www.youtube.com/watch?v=HIak2kthBWQ

#SD2 #Galactica #DigiHavel #StableDiffusion #METAAI

Originally published on Facebook — link to post